ProofGate: a merge queue that proves agent-written code safe before it ships
The problem
Agent-written code is flooding pull requests, and the AI reviewers teams already pay for can only pattern-match: they catch style issues and likely bugs, but cannot prove a security property. A Meta engineer's viral post about shipping code nobody understands drew thousands of comments from engineers in the same spot, and one documented case had an AI-generated password-hashing function with a salt flaw that only surfaced in an audit months later.
The idea
Why now
Three capabilities landed within days of each other: AlphaProof Code showed agents formally verifying Signal's real cryptographic code in four weeks (10 November); decision models like Strands Decider answer in tens of milliseconds, making per-PR triage nearly free; and LiveReview proved teams want risk-aware auto-approval instead of reviewing everything by hand.
What it combines
AlphaProof Code's formal verification of real code plus Strands Decider's near-free routing decisions plus LiveReview's risk-aware auto-approval. Verification without triage is too slow and expensive to run on every PR; triage without verification is just smarter linting; auto-approval without either is a gamble. Together they make a queue that is both fast and provably safe on the paths that matter.
MVP
Weekend MVP: a GitHub Action that runs on PRs touching a configurable list of sensitive paths, calls a decision model to classify the diff, and for crypto/auth files runs a small proof check on one function, posting the result as a check. Skip: the full queue, auto-merge, and multi-repo proofs.
How it makes money
The moat
The proof corpus is a per-repo data advantage: every verified path becomes a reusable proof artifact, so the queue gets faster and cheaper on each repo the longer it runs — a new entrant starts at zero proofs. GitHub's own review tooling is velocity-oriented and would have to rebuild around proof generation and take on liability for 'verified' claims; pattern-matching incumbents cannot bolt formal verification onto a suggester in one quarter.
Distribution
GitHub App sold to engineering-platform and AppSec teams; the wedge is security teams at companies running AI-coding mandates, where the volume of agent PRs has already outrun human review. B2B, per-developer seats like the review tools it replaces.
Why it wins
CodeRabbit, Greptile and Qodo pattern-match and suggest fixes; ProofGate proves. The deliverable is not a comment thread but a mathematical guarantee on the security-critical paths, attached to the merge as an artifact a security team can audit.
Risks
Formal verification is slow and can fail on real-world code with side effects, so proof coverage may stay narrow. The MVP de-risks it by scoping to one function class and measuring the proof success rate before promising queue-wide coverage.
Build it with
- AlphaProof Code: AI agents formally verify Signal's crypto code in four weeksProves security properties of real crypto code — the verification engine the queue gates on.
- Strands Decider: AWS's open-weight decision model that picks instead of generatingNear-free per-PR triage: which diffs need verification and which can auto-approve.
- LiveReview: risk-aware AI code review that auto-approves safe PRsThe risk-aware auto-approval pattern for the safe majority of PRs.
Repo to start from
proofgate: a GitHub Action and merge queue that routes security-critical agent PRs to formal verification and auto-approves the rest.
Evidence
- CodeRabbit pricing: $24-72 per developer per month
- Ask HN: How to deal with shipping AI-written code you don't understand?
- Strands Decider: open-weight decision model
- LiveReview: risk-aware AI code review
Get the week's best AI launches, plus 3 ideas worth building
One email every Saturday. Ranked by traction, not hype. Free.